A method of inducing a top-down hierarchical categorizer includes providing a set of labeled training items. Each labeled training item includes an associated label representing a single category assignment for the training item. A set of unlabeled training items is provided. A prior is associated with the set of unlabeled training items that is independent of any particular feature contained in the unlabeled training items. The prior represents a plurality of possible category assignments for the set of unlabeled training items. A top-down hierarchical categorizer is induced with a machine learning algorithm based on the set of labeled training items, the set of unlabeled training items, and the prior.
|
1. A computer-implemented method for inducing a top-down hierarchical categorizer, comprising:
providing a set of labeled training items, each labeled training item including an associated label representing a single category assignment for the training item;
providing a set of unlabeled training items;
associating a prior with the set of unlabeled training items that is independent of any particular feature contained in the unlabeled training items, the prior representing a plurality of possible category assignments for the set of unlabeled training items; and
inducing a top-down hierarchical categorizer with a machine learning algorithm based on the set of labeled training items, the set of unlabeled training items, and the prior.
22. A computer-readable medium having computer-executable instructions for performing a method of inducing a top-down hierarchical categorizer based on labeled training items and unlabeled training items, each labeled training item including a label identifying a single category assignment, each unlabeled training item including prior information representing a plurality of possible category assignments for the unlabeled training item, the method comprising:
providing a plurality of categories organized in a hierarchy of categories; inducing a plurality of categorizers corresponding to the plurality of categories; and
associating a set of features with each category based on features of labeled training items assigned to that category and based on features of unlabeled training items with prior information representing category assignments that map to that category.
15. A computer-implemented system for inducing a top-down hierarchical categorizer from a set of labeled training items and a set of unlabeled training items, each labeled training item including an associated label representing a single category assignment for the training item, the system comprising:
means for associating a prior with each unlabeled training item that is independent of any particular feature contained in the unlabeled training item, the prior representing a plurality of possible category assignments for the unlabeled training item; and
means for inducing a plurality of categorizers based on a machine learning algorithm, the set of labeled training items, the set of unlabeled training items, and the priors associated with the unlabeled training items, the plurality of categorizers including a root categorizer and a plurality of sub-categorizers organized in a tree structure.
8. A computer-implemented system for inducing a top-down hierarchical categorizer based on labeled training items and unlabeled training items, each labeled training item including a plurality of features and a label identifying a single category assignment, each unlabeled training item including a plurality of features and prior information representing a plurality of possible category assignments for the unlabeled training item, the system comprising:
a plurality of categories organized in a hierarchy of categories; and an inducer for inducing a plurality of categorizers corresponding to the plurality of categories, the inducer configured to induce each categorizer based on the features of labeled training items assigned to categories under that categorizer and based on the features of unlabeled training items with prior information representing category assignments that map to a category under that categorizer.
2. The method of
determining whether the possible category assignments for the set of unlabeled training items all map to a single category; and
using the sot of unlabeled training items as training examples for the single category.
3. The method of
4. The method of
5. The method of
6. The method of
7. The method of
9. The system of
10. The system of
11. The system of
12. The system of
13. The system of
14. The system of
16. The system of
means for determining whether the possible category assignments for a first prior all map to a common category; and
using the training item associated with the first prior as a training example for the common category.
17. The system of
18. The system of
19. The system of
20. The system of
21. The system of
23. The computer-readable medium of
using the unlabeled training items as positive training examples for at least one of the categories in the hierarchy of categories.
24. The computer-readable medium of
using the unlabeled training items as negative training examples for at least one of the categories in the hierarchy of categories.
25. The computer-readable medium of
26. The computer-readable medium of
27. The computer-readable medium of
inducing a plurality of categorizers corresponding to the plurality of categories based on a general-purpose induction algorithm.
28. The computer-readable medium of
29. The computer-readable medium of
restructuring the hierarchy of categories based on the prior information.
|
The present invention relates to categorization systems and more particularly to a system and method for inducing a top-down hierarchical categorizer.
Categorization involves assigning items (e.g., documents, products, patients, etc.) into categories based on features of the items (e.g., which words appear in a document), and possibly subject to a degree of confidence. For example: vehicle X that has the features
number of seats = 55
color = yellow
belongs to the category “school buses” with probability 95%.
Hierarchical categorization is the problem of categorizing where the categories are organized in a hierarchy. The field's terminology has a number of common synonyms, such as:
categorization = classification, prediction
features = attributes, properties
categories = classes, subtopics
confidence = degree of belief, certainty
items = cases, examples
machine learning = supervised learning, induction
In the past, many different systems have been developed for categorizing different types of items. The earliest systems used manual assignment of documents to categories, for example, by human experts. This is currently the dominant method, which is used in libraries, as well as by popular Internet search engine companies.
Disadvantages of manual assignment include the fact that it requires a large amount of human resources and it is labor-intensive. In addition, manual assignment is somewhat error-prone and may lead to inconsistencies if people are assigning documents to categories based on different criteria, different interpretations of criteria, or different levels of expertise.
To be less subjective, rule-based assignment of documents to categories, including rules based on keywords, has been developed for use with computer systems. This approach uses rules such as “if the document contains the words ‘football’, and ‘goal’, and ‘umpire’ and not the word ‘national’ then assign it to the category ‘local football.’”
Mostly, human domain experts author these rules, possibly with the aid of keyword identification tools (such as word counters). These rules usually are comprised of Boolean combinations of keyword occurrences (possibly modified by counts such as “if the term ‘national’ is used at least 5 times then assign to ‘national baseball’”). These rules can be executed automatically, so this solution can be used to automatically assign documents to categories. Examples of human-authored rule classifier systems include a topics search engine by Verity Corp., and email routing software by Kana Communications Inc.
The disadvantages of rule-based assignment are that the accuracy of these rules is often very poor. Depending on the authoring of the rules, either the same document is assigned to many categories, including many wrong categories, or to too few categories, in which case documents do not appear in the categories they should. Another disadvantage is that the rules are difficult to author and maintain, and the interaction of the rules (so-called “chaining”) is difficult to understand (and debug), so that unexpected assignments of documents to categories may occur.
Categorizers may be built manually by people authoring rules/heuristics, or else built automatically via machine learning, wherein categorizers are induced based on a large training set of items. Each item in the training set is typically labeled with its correct category assignment. The use of predefined categories implies a supervised learning approach to categorization, where already-categorized items are used as training data to build a model for categorizing new items. Appropriate labels can then be assigned automatically by the model to new, unlabeled items depending on which category they fall into. Typically, the larger the training set, the better the categorization accuracy. However, it typically costs something (e.g., human labeling effort) to prepare the training set.
Examples of machine learning algorithms include the well-known Naïve Bayes and C4.5 algorithms, support vector machines, and commercial offerings such as those of Autonomy Inc., and Moho Mine Inc.
One type of categorizer that can be induced by such machine learning algorithms is a top-down hierarchical categorizer (also referred to as a Pachinko classifier). A top-down hierarchical categorizer typically considers a topic hierarchy one level at a time. At each level, there are typically one or more categorizers that, when assigned a document, pick a category at the next level based on features of the document.
A major barrier to using machine-learning categorization technology is that it requires a significant amount of training data, the gathering of which involves significant costs, delays and/or human labor.
One form of the present invention provides a method of inducing a top-down hierarchical categorizer. A set of labeled training items is provided. Each labeled training item includes an associated label representing a single category assignment for the training item. A set of unlabeled training items is provided. A prior is associated with the set of unlabeled training items that is independent of any particular feature contained in the unlabeled training items. The prior represents a plurality of possible category assignments for the set of unlabeled training items. A top-down hierarchical categorizer is induced with a machine learning algorithm based on the set of labeled training items, the set of unlabeled training items, and the prior.
In the following detailed description of the preferred embodiments, reference is made to the accompanying drawings that form a part hereof, and in which is shown by way of illustration specific embodiments in which the invention may be practiced. It is to be understood that other embodiments may be utilized and structural or logical changes may be made without departing from the scope of the present invention. The following detailed description, therefore, is not to be taken in a limiting sense, and the scope of the present invention is defined by the appended claims.
One type of categorizer 106 that may be induced by machine learning algorithms is a top-down hierarchical categorizer. As shown in
Processor 302 is coupled to display interface 312, main memory 316, secondary memory 318, and input device 320 via bus 310. Display interface 312 is coupled to display 314. Main memory 316 stores data, and application program instructions for execution by processor 302. In one embodiment, secondary memory 318 includes a disk drive, CD-ROM drive, and/or other non-volatile storage systems. An input device 320, such as a keyboard, allows a user to enter data and otherwise interact with system 300.
As shown in
It will be understood by a person of ordinary skill in the art that functions performed by system 300 may be implemented in hardware, software, firmware, or any combination thereof. The implementation may be via a microprocessor, programmable logic device, or state machine. Components of embodiments of the present invention may reside in software on one or more computer-readable mediums. The term computer-readable medium as used herein is defined to include any kind of memory, volatile or non-volatile, such as floppy disks, hard disks, CD-ROMs, flash memory, read-only memory (ROM), and random access memory.
In one embodiment, items to be categorized by top-down hierarchical categorizer 308 comprise documents, such as electronic documents accessible on the Internet. However, it will be understood that further embodiments of the invention are applicable to other types of items and other network architectures (e.g., client/server, local, intermediate or wide area networks), dedicated database environments, or other configurations.
Training items 402 include training items with labels 402A and training items with priors 402B. In one embodiment, training items 402 are simply raw documents that do not include associated labels or priors, and labels and priors are associated with the documents after the documents have been featurized. In one embodiment, featurizer 404 produces a feature vector for each training item in training items 402. The set of feature vectors generated for training items 402 is referred to as training feature vectors 406.
The training feature vectors 406 are provided to training set generator 410. Training set generator 410 also receives hierarchy of categories 408, labels 403, and prior information 413. Based on these inputs, training set generator 410 generates a training set 412 that includes a plurality of feature vectors, with each feature vector having an associated label 403 or prior information 413.
According to one embodiment of the present invention, the hierarchy of categories 408 is provided by a user. The hierarchy 408 represents a “tree” of categories that has “branches” that end at “leaves”. The leaves are the places in the hierarchy where there are no further subdivisions under a given category.
In one embodiment, labels 403 are implemented as a table that specifies category assignments for training items 402A (e.g., document 1023812 is assigned to category A12, document 1023813 is assigned to category B11, etc.), and each training item in training items 402A has a single category from the hierarchy of categories 408 associated with the item (i.e., the item's “label” 403).
In one embodiment, prior information 413 is implemented as one or more tables that specify for each training item in training items 402B a set of possible categories from the hierarchy of categories 408 for the item (e.g., document 1023814 is assigned to one of categories A11 or A12, document 1023815 is assigned to one of categories A11, A21, or A22, etc.). One embodiment of prior information 413 is illustrated in
In one form of the invention, inducer 306 repeatedly calls induction algorithm 416 and training set generator 410 as represented by the dotted arrows in
In one embodiment, system 400 is implemented as one or more software modules executable by computer system 300 (shown in
During categorization according to one embodiment, hierarchical categorizer 308 works top-down. Top categorizer 500 chooses which of the two branch paths below it to follow, and each sub-categorizer 502 selects which local branch path below that sub-categorizer 502 to follow. Further sub-categorizers 502 would be added for more complex categorization systems. Each sub-categorizer 502 represents the beginning of a “subtree.” The hierarchical categorizer 308 shown in
In one embodiment, when training each sub-categorizer 502, all of the training items with labels 402A that fall within the subtree of that sub-categorizer 502 have their labels temporarily mapped to the (smaller) set of branch choices directly below that sub-categorizer 502. For example, during training of sub-categorizer 502A, training items 402A having labels corresponding to category 504A-11 or 504A-12 would be temporarily mapped to category 504A-1, and would be training examples for category 504A-1. Similarly, training items 402A having labels corresponding to category 504A-21 or 504A-22 would be temporarily mapped to category 504A-2. The induction algorithm 416 looks at the features of each training item and the category that the training item belongs in, and essentially determines what the pattern of features is that puts the training item in that category.
In one form of the invention, a “prior” for a document indicates that the document's correct categorization falls somewhere within a certain set of categories, whereas the document's label names a single correct category assignment for the document.
For example, assume that books are being categorized into dozens of categories. Assume that category 504A in
In one embodiment, prior information in tables 600 and 610 is accessed and used during training of hierarchical categorizer 308. Table 600 is accessed to determine if a document (represented by its document ID number 602) includes an associated prior identifier 604. If the document does have an associated prior identifier 604, table 610 is accessed to identify the prior 614 associated with the prior identifier 604. The identified prior 614 indicates possible categories 504 that the document under consideration might fall under. For example, the first prior 614 listed in table 610 is “A11 or A22” which represents category 504A-11 or 504A-22 in
In one embodiment, prior information is manually entered and associated with training documents as the documents are entered into the system for training. This typically involves far less effort than assigning a specific label to each case. For example, if a set of 15,000 documents are received from Company X, and it is known that documents from Company X all have a prior identifier 604 of “1,” then a human operator can easily associate this prior information with all 15,000 documents at one time. In contrast, to go through each one of the 15,000 documents individually and assign a specific label would take a much larger amount of time and human resources.
For the purposes of obtaining the most leverage from the priors 614, in one embodiment, the hierarchy 408 is preferably organized in such a way that most training items' priors 614 have all of their possible categories 504 falling within a single subtree of the hierarchy 408, as small as possible. For example, if many training items 402B come with a prior 614 that specifies various non-fiction categories 504, the non-fiction categories 504 would preferably be gathered into one small subtree (e.g., the subtree beginning at category 504B-2), rather than having these categories 504 spread randomly throughout the hierarchy 408.
Priors 614 are as valuable as labels whenever all of their permissible categories 504 map to the same branch choice. For example, in the tree shown in
In one embodiment, the hierarchy of categories 408 may be restructured automatically or by hand for the purpose of yielding greater leverage of the priors 614. For example, if a prior 614 for a very large number of training items 402B specifies “A12 or A22”, then the hierarchy 408 might be profitably permuted by placing these two nodes as siblings under a replacement node A1′ and placing A11 and A21 into a replacement node A2′. Furthermore, portions of the hierarchy 408 might be flattened to leverage a prior 614. In this example, A1 and A2 might be eliminated to make A11, A12, A21, and A22 siblings of A3 directly under A. Items within this prior 614 may then be used as negative training cases for A11, A21 and A3. In the limit, this strategy may also be used to completely flatten the hierarchy 408 to a single root node containing all the categories as direct children.
In one embodiment, priors 614 are phrased as constraints on categories, and are used as positive or as negative training examples. For example, if a training item 402B has prior information that indicates that it is in 504A-12 or 504A-22, in one embodiment, the training item is used in the sub-categorizer 502A as a training example of “not 504A-3”.
In one embodiment, priors are phrased as probabilistic constraints or information. For example, it can be specified that there is an 80% chance that item X belongs in category 504A-1 and a 20% chance that it is possibly in category 504A-3. Depending on the application and the desired set-up that is chosen, the probabilities can be required to sum up to one, indicating mutually exclusive and exhaustive categories, or more flexible category information can be used.
In one form of the invention, probabilistic information is incorporated by assigning a weight to each training item 402 that is proportional to the probability of the training item 402 being assigned to a particular category 504. For instance, a “real” labeled training item 402A for category 504A-1 gets weight 1, whereas an example that has an 80% chance of being in 504A-1 gets counted with weight 0.8.
By using probabilistic information, in one embodiment, prior information may also be used even if the categories 504 to which a training item belongs (according to the prior information) extend beyond one sub-tree. In one form of the invention, this is done by applying a threshold (e.g., if the probability “mass” exceeds 75% then count it as an example) or by using a weight as described above, or using some similar method such as using a so-called lambda vector (i.e., the possible categories for a prior are represented by fuzzy set membership information). For example, suppose a prior 614 gives a 45% probability mass to category 504A-11, 45% to category 504A-12, and 10% to category 504B-11. In one form of the invention, training items under this prior 614 are used as positive examples for category 504A-1 when training the sub-categorizer 502A, either because 90% is deemed to be “good enough” (exceeds a specified threshold), or a weight of 0.9 exceeds a specified threshold, or some similar threshold is satisfied.
In one embodiment, probabilistic information is obtained by running a large number of training items 402B with a particular prior 614 through an existing categorizer and observing the probability distribution. For example, assume that a collection of documents is received from vendor X for a document hierarchy. Using a categorizer, a distribution can be derived as to where the items in this collection belong (i.e., 32% of items from vendor X are categorized in category 504A-11). A new categorizer is then trained using this additional prior information.
The following is a pseudo-code procedure (“Train”) for implementing one embodiment of inducer 306 and training set generator 410, for inducing a top-down hierarchical categorizer 308 according to one form of the present invention:
Pseudo Code Example
1
procedure Train(H,T,P,A): HC
2
3
Let S={S1, . . . ,SN} be the direct children of the root of H.
4
5
Let T2 be a copy of T, where the labels have been restricted to
6
the set S.
7
8
Add to T2 any training case in P where its set of allowable
9
labels falls entirely under a single node SJ.
10
11
Trivial Base Case
12
IF H contains only a single category,
13
THEN return a categorizer that always returns this single
14
category, no matter what its input feature vector.
15
(ELSE continue this procedure)
16
17
Induction Case
18
Apply the induction algorithm A to the items in the training
19
set T2 to produce a new categorizer C.
20
21
Recursion
22
for each subtree s in S {
23
Prepare Hs', Ts', and Ps' which are each restricted to the
24
subtree s.
25
Recursively call Train(Hs', Ts', Rs', A) and store the
26
result Cs for use below
27
}
28
29
return a hierarchical categorizer (HC) that
30
(1) applies the categorizer C and determines which subtree
31
s is most appropriate; and
32
(2) applies the categorizer Cs for the chosen subtree and
33
returns its result.
There are four inputs to the procedure Train: (1) H—a hierarchy of categories 408; (2) T—a training set (i.e., a set of feature vectors 406 that are labeled with their correct category assignment 504); (3) P—a prior training set (i.e., like the training set T, except the label information does not pinpoint a single category 504, but instead a subset of categories 504 where the feature vector 406 may belong); and (4) A—a general-purpose induction algorithm 416 (e.g., C4.5, Naïve Bayes, etc.) that, given an input training set 412, outputs a categorizer function C.
In one embodiment, parameter A is actually a pointer to a function that implements the general-purpose induction algorithm 416. In another embodiment, the algorithm A is not a parameter, but is hard-coded into the training routine. The procedure Train outputs HC, which is a trained hierarchical categorizer 308.
The procedure Train is recursive (i.e., the procedure calls itself, but typically with different parameters for each call). The initial call to the procedure Train passes input variables H, T, and P that contain information that is relevant for the whole hierarchy of categories 408 (i.e., H represents the entire hierarchy 408, and T and P represent the entire training set 412). Subsequently, in one form of the invention, when the procedure Train calls itself, the input variables H, T and P contain only the information that is pertinent to the subtree of the hierarchy 408 that the procedure is dealing with at that moment, as described in further detail below.
At line 3 of the procedure Train, the statement “Let S={S1, . . . , SN} be the direct children of the root of H” indicates that the set of top-level choices in the hierarchy 408 are assigned to S1, S2, . . . , SN. For the example hierarchy shown in
At lines 5–6 of the procedure Train, the statement “Let T2 be a copy of T, where the labels have been restricted to the set S” indicates that T2 represents a transformed version of the training set 412, where all of the labels for categories 504 below the set S (i.e., below categories 504A and 504B) are re-mapped to labels for their ancestor categories 504 within the set S. For example, if an item were labeled S2/1/2, which corresponds to category 504B-12 in the embodiment of
At lines 8–9 of the procedure Train is the statement “Add to T2 any training case in P where its set of allowable labels falls entirely under a single node Sj.” For the first call to the procedure Train, the set S includes S1 (i.e., category 504A) and S2 (i.e., category 504B). Thus, any training case in P that has a set of allowable labels that falls entirely under either the subtree at category 504A or the subtree at category 504B is added to T2. Such cases are as good as labeled training cases for S1 and S2. In one embodiment, items whose set of allowable labels falls under multiple nodes Sj are excluded from T2. In this embodiment, if a training case in P has a set of allowable labels that falls somewhere under both categories 504A and 504B, the training case is excluded from T2. For example, if a prior 614 indicates that the training item could be in category 504A-1, category 504B-11, or category 504B-12, that training item is not useful in deciding between category 504A or 504B and is excluded.
Lines 11–15 of the procedure Train handle a trivial base case. As indicated therein, if the hierarchy of categories, H, contains a single category 504, then the procedure returns a trivial categorizer that always returns this single category 504, no matter what its input feature vector. If H contains more than one category 504, then the procedure continues.
Lines 17–19 of the procedure Train handle an induction case. As indicated therein, the induction algorithm A is applied to the items in the training set T2 to produce a new categorizer C. If all the items in T2 fall under a single branch, then the induction algorithm A need not be run and C is the trivial classifier that selects the single populated branch. The new categorizer will be used at categorization time to select the next branch in the top-down decision process. For the example illustrated in
Lines 21–26 of the procedure Train specify a recursion process. For each subtree s in S, the procedure prepares Hs′, Ts′, and Ps′ which are each restricted to the subtree s, recursively calls itself with the inputs Hs′, Ts′, Rs′, and A, and stores a resulting sub-categorizer, Cs. In one embodiment, any training item in P that has a set of allowable labels that falls into multiple subtrees s is excluded from P′. For the embodiment illustrated in
At lines 29–33 of the procedure Train, a hierarchical categorizer (HC) 308 is returned. In one embodiment, the hierarchical categorizer 308 does the following: (1) applies the categorizer C and determines which subtree s is most appropriate; and (2) applies the categorizer Cs for the chosen subtree and returns its result.
In one embodiment, additional recursive calls to the procedure Train are made to induce further sub-categorizers 502 for additional levels in the hierarchy (e.g., sub-categorizers 502A-1, 502A-2, 502A-3, 502B-1, and 502B-2).
The above pseudo code example for the procedure Train is provided to illustrate one embodiment of inducer 306 and training set generator 410. It will be readily apparent to persons of ordinary skill in the art that various modifications and additions may be made for alternative embodiments. As one example, rather than passing a parameter A (i.e., a pointer to the general purpose induction algorithm 416) to the procedure Train, in an alternative embodiment, the code for the induction algorithm 416 is included in a subroutine.
After top-down hierarchical categorizer 308 has been trained, it is ready to classify new, unlabeled items.
The categorization process starts with an unclassified item 702 that is to be categorized, such as a raw document. The unclassified item 702 is provided to featurizer 704. Featurizer 704 extracts features from the unclassified item 702, such as whether a word 1 was present and a word 2 was absent, or the word 1 occurred five times and the word 2 did not occur at all. The features from the featurizer 704 are used to create a list of features 706. The list of features 706 is provided to hierarchical categorizer 308, which selects the best category 504 for item 702 based on the provided list of features 706.
During training of hierarchical categorizer 308 according to one embodiment, a set of features is associated with each of the categories 504. The term “level of goodness” is used herein to describe how good the fit is between the list of features 706 of a document 702 to be categorized and the previously determined features of a category 504. There are many different ways of determining level of goodness, such as Naïve Bayes, C4.5, Bayesian networks, rule-based multi-category categorizers that output some level of goodness, conditional probability statements, or simple heuristics, among others, or a combination of the foregoing.
In one embodiment, root categorizer 500 of hierarchical categorizer 308 computes for item 702 a level of goodness of the match between the item 702 and the categories 504A and 504B, and then applies a decision criterion for determining whether the level of goodness is high enough to assign the item 702 to one of these categories 504A or 504B. Similarly, each sub-categorizer 502 that is assigned the item 702 computes a level of goodness of the match between the item 702 and the categories 504 under that sub-categorizer 502, and applies a decision criterion for determining whether the degree of goodness is high enough to assign the document to one of these categories 504. This process is repeated deeper into the hierarchy until a final, correct category 504 is identified.
Priors have been used before in a couple of different categorization techniques. In a first technique, the overall categorization problem has been broken up into sub-problems based on a prior, and then machine learning has separately been applied to each sub-problem. The machine learning algorithm itself does not use the prior information, but rather relies on labeled training items for each of the sub-problems. This technique effectively makes a composite categorizer that depends foremost on the prior attribute.
As a potential example of this technique, consider the many Internet directories, such as Yahoo (http://www.yahoo.com) and Infoseek (http://www.infoseek.com) that are largely manually organized in preset hierarchies. For shopping web sites, such as Yahoo shopping, there can be thousands or hundreds of thousands of different products from thousands of different stores. For such sites, rather than manually placing each product into a category within the hierarchy, people typically write textual queries to sort out the various products and identify categories for each product. To further simplify the categorization process, the web site may require companies to identify a top-level category that the company's products fall under. For example, for the hierarchy shown in
This first technique has several disadvantages. One disadvantage is that the technique depends on the priors being disjoint, which is not always the case. For example, some book publishers produce only non-fiction, and other publishers produce only non-fiction technical material (a smaller set of topics under the non-fiction topic). Thus, priors having different granularities may be provided, but the more specific prior information is not leveraged in the first technique. The additional information that the non-fiction material produced by certain publishers is always technical material is not used in the first technique.
A second disadvantage of the first technique is that there is often some non-zero cost associated with obtaining a prior, which one is typically willing to pay for training, but is not typically willing to pay for every future use of the trained categorizer. In other words, using the first technique, even after the categorizer has been trained, priors must be obtained for each and every item to be categorized, so that it can be assigned to the appropriate branch in the hierarchy for further categorization. Obtaining these priors has an associated cost.
A third disadvantage of the first technique is that priors may not be available on many incoming cases to be categorized.
A second categorization technique that uses priors treats a prior as a predictive feature for every case. Thus, for each training case, the prior is included as an additional feature. This technique carries the second and the third disadvantages described above. Regarding the second disadvantage, since the prior is used as a predictive feature, it is important to have the prior at categorization time. And it may be costly to continually have to obtain this prior information for every case that needs to be categorized in the future. Regarding the third disadvantage, if the prior is not available at categorization time, the prior could be treated as a missing value. However, the induced categorizer might have learned to depend heavily on that attribute, and accuracy will suffer if the missing attribute is treated with standard missing-value techniques (e.g., filling it in with the majority value or most likely value given the other attributes).
Embodiments of the present invention do not have the disadvantages associated with the above-described previous techniques for using prior information. In one embodiment, prior information need not be supplied for every training case. One embodiment of the present invention uses priors essentially as labels during training, and then the prior information is not needed during later categorization of unlabeled documents. In one embodiment, priors are used during training of a hierarchical top-down categorizer to increase the amount of training information for the categorizer, and thereby increase the categorization accuracy of the categorizer with relatively minor added cost. In many situations, various priors are freely available, but labels are costlier to obtain. In one embodiment, the use of prior information in training a top-down hierarchical categorization system reduces the amount of labeled training data that is needed for the system. In one embodiment, if the priors are available during classification (not just at training), they may be used to restrict the eligible categories for classification, eliminating some erroneous categories from consideration.
Although specific embodiments have been illustrated and described herein for purposes of description of the preferred embodiment, it will be appreciated by those of ordinary skill in the art that a wide variety of alternate and/or equivalent implementations may be substituted for the specific embodiments shown and described without departing from the scope of the present invention. Those with skill in the chemical, mechanical, electro-mechanical, electrical, and computer arts will readily appreciate that the present invention may be implemented in a very wide variety of embodiments. This application is intended to cover any adaptations or variations of the preferred embodiments discussed herein. Therefore, it is manifestly intended that this invention be limited only by the claims and the equivalents thereof.
Forman, George H., Suermondt, Henri Jacques
Patent | Priority | Assignee | Title |
11775588, | Dec 24 2019 | CIGNA INTELLECTUAL PROPERTY, INC | Methods for providing users with access to data using adaptable taxonomies and guided flows |
7593904, | Jun 30 2005 | MICRO FOCUS LLC | Effecting action to address an issue associated with a category based on information that enables ranking of categories |
7711779, | Jun 20 2003 | Microsoft Technology Licensing, LLC | Prevention of outgoing spam |
7756799, | Oct 27 2006 | MICRO FOCUS LLC | Feature selection based on partial ordered set of classifiers |
7930353, | Jul 29 2005 | Microsoft Technology Licensing, LLC | Trees of classifiers for detecting email spam |
8046832, | Jun 26 2002 | Microsoft Technology Licensing, LLC | Spam detector with challenges |
8065370, | Nov 03 2005 | Microsoft Technology Licensing, LLC | Proofs to filter spam |
8095548, | Oct 14 2008 | Saudi Arabian Oil Company | Methods, program product, and system of data management having container approximation indexing |
8224905, | Dec 06 2006 | Microsoft Technology Licensing, LLC | Spam filtration utilizing sender activity data |
8494998, | May 04 2007 | Microsoft Technology Licensing, LLC | Link spam detection using smooth classification function |
8719073, | Aug 25 2005 | Hewlett-Packard Development Company, L.P. | Producing a measure regarding cases associated with an issue after one or more events have occurred |
8805754, | May 04 2007 | Microsoft Technology Licensing, LLC | Link spam detection using smooth classification function |
8818838, | Mar 12 2009 | Henry Rak Consulting Partners, LLC | System and method for efficiently developing a hypothesis regarding the structure of a market |
9092422, | Dec 30 2009 | GOOGLE LLC | Category-sensitive ranking for text |
9116641, | Nov 30 2004 | Panduit Corp | Market-based labeling system and method |
9213767, | Aug 10 2009 | MICRO FOCUS LLC | Method and system for characterizing web content |
Patent | Priority | Assignee | Title |
5627980, | Mar 20 1995 | Xerox Corporation | Elision based presentation of ordered data |
5694524, | Feb 15 1994 | R R DONNELLEY & SONS COMPANY | System and method for identifying conditions leading to a particular result in a multi-variant system |
5781698, | Oct 31 1995 | Carnegie Mellon University | Method of autonomous machine learning |
5796863, | Aug 08 1995 | Apple Inc | Method for training an adaptive statistical classifier to balance unigram prior factors |
5930803, | Apr 30 1997 | RPX Corporation | Method, system, and computer program product for visualizing an evidence classifier |
5999893, | May 02 1997 | NAVY, THE UNITED STATES OF AMERICA AS REPRESENTED BY THE SECRETARY OF THE | Classification system and method using combined information testing |
6006225, | Jun 15 1998 | Amazon Technologies, Inc | Refining search queries by the suggestion of correlated terms from prior searches |
6026399, | May 30 1997 | RPX Corporation | System and method for selection of important attributes |
6178416, | Jun 15 1998 | PARKER, JAMES U | Method and apparatus for knowledgebase searching |
6182058, | Feb 28 1997 | RPX Corporation | Bayes rule based and decision tree hybrid classifier |
6233575, | Jun 24 1997 | International Business Machines Corporation | Multilevel taxonomy based on features derived from training documents classification using fisher values as discrimination values |
6430558, | Aug 02 1999 | TUMBLEWEED HOLDINGS LLC | Apparatus and methods for collaboratively searching knowledge databases |
6489968, | Nov 18 1999 | A9 COM, INC | System and method for exposing popular categories of browse tree |
6516308, | May 10 2000 | AT&T Corp | Method and apparatus for extracting data from data sources on a network |
Executed on | Assignor | Assignee | Conveyance | Frame | Reel | Doc |
Jul 30 2002 | FORMAN, GEORGE HENRY | Hewlett-Packard Company | ASSIGNMENT OF ASSIGNORS INTEREST SEE DOCUMENT FOR DETAILS | 013639 | /0838 | |
Jul 31 2002 | SUEMONDT, HENRI JACQUES | Hewlett-Packard Company | ASSIGNMENT OF ASSIGNORS INTEREST SEE DOCUMENT FOR DETAILS | 013639 | /0838 | |
Aug 02 2002 | Hewlett-Packard Development Company, L.P. | (assignment on the face of the patent) | / | |||
Jan 31 2003 | Hewlett-Packard Company | HEWLETT-PACKARD DEVELOPMENT COMPANY, L P | ASSIGNMENT OF ASSIGNORS INTEREST SEE DOCUMENT FOR DETAILS | 013776 | /0928 |
Date | Maintenance Fee Events |
Jul 24 2009 | M1551: Payment of Maintenance Fee, 4th Year, Large Entity. |
Sep 06 2013 | REM: Maintenance Fee Reminder Mailed. |
Jan 24 2014 | EXP: Patent Expired for Failure to Pay Maintenance Fees. |
Feb 24 2014 | EXP: Patent Expired for Failure to Pay Maintenance Fees. |
Date | Maintenance Schedule |
Jan 24 2009 | 4 years fee payment window open |
Jul 24 2009 | 6 months grace period start (w surcharge) |
Jan 24 2010 | patent expiry (for year 4) |
Jan 24 2012 | 2 years to revive unintentionally abandoned end. (for year 4) |
Jan 24 2013 | 8 years fee payment window open |
Jul 24 2013 | 6 months grace period start (w surcharge) |
Jan 24 2014 | patent expiry (for year 8) |
Jan 24 2016 | 2 years to revive unintentionally abandoned end. (for year 8) |
Jan 24 2017 | 12 years fee payment window open |
Jul 24 2017 | 6 months grace period start (w surcharge) |
Jan 24 2018 | patent expiry (for year 12) |
Jan 24 2020 | 2 years to revive unintentionally abandoned end. (for year 12) |